myBit-Llama2-jp-127M-2
This model was built by referring to the config in TinyLlama/TinyLlama-1.1B-Chat-v1.0, 123M The model is a pre-trained Bit-Llama2 of Parameters with only 1 epoch on a Japanese dataset. The dataset used is range3/wiki40b-ja.
- Loss: 3.0972
Model description
Github： BitNet-b158
More information about this model can be found in the following pages:
BitNet&BitNet b158の実装
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.00024
- train_batch_size: 96
- eval_batch_size: 96
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: polynomial
- lr_scheduler_warmup_steps: 5000
- num_epochs: 1
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 6.9469 | 0.05 | 2000 | 4.7495 |
| 4.4263 | 0.1 | 4000 | 4.2282 |
| 4.0799 | 0.15 | 6000 | 3.9351 |
| 3.8199 | 0.2 | 8000 | 3.7209 |
| 3.6462 | 0.25 | 10000 | 3.5755 |
| 3.5239 | 0.29 | 12000 | 3.4803 |
| 3.4727 | 0.34 | 14000 | 3.4181 |
| 3.3953 | 0.39 | 16000 | 3.3752 |
| 3.3562 | 0.44 | 18000 | 3.3395 |
| 3.3272 | 0.49 | 20000 | 3.3166 |
| 3.2965 | 0.54 | 22000 | 3.2869 |
| 3.2771 | 0.59 | 24000 | 3.2680 |
| 3.2545 | 0.64 | 26000 | 3.2478 |
| 3.235 | 0.69 | 28000 | 3.2276 |
| 3.2189 | 0.74 | 30000 | 3.2054 |
| 3.1973 | 0.79 | 32000 | 3.1910 |
| 3.1793 | 0.83 | 34000 | 3.1675 |
| 3.1572 | 0.88 | 36000 | 3.1487 |
| 3.1406 | 0.93 | 38000 | 3.1279 |
| 3.1144 | 0.98 | 40000 | 3.0972 |
Framework versions
- Transformers 4.38.2
- Pytorch 2.1.0+cu121
- Datasets 2.18.0
- Tokenizers 0.15.2